Recover Dataproc jobs after worker crashes by 1fanwang · Pull Request #72417 · apache/airflow

1fanwang · 2026-09-02T05:45:33Z

A synchronous Dataproc job can outlive its worker, but a task retry currently submits a second job instead of reconnecting. After this change, the retry restores the API-returned job ID from task state and resumes waiting for the same job.

DataprocSubmitJobOperator uses the AIP-103 resumable-job contract only on its synchronous wait path. Active jobs reconnect, completed jobs return without another submission, and terminal or missing jobs submit fresh. Recovery restores the operator job ID and extra link state used by cancellation. Asynchronous, deferrable, direct-to-trigger, OpenLineage, and pre-Airflow 3.3 behavior stay unchanged.

An explicit request_id is still passed unchanged. Dataproc may therefore return the prior terminal job when a retry resubmits with the same idempotency key.

Testing

$ PATH="$PWD/.venv/bin:$PATH" AIRFLOW_HOME="$PWD/.build/airflow-home-dataproc-final" uv run --project providers/google pytest providers/google/tests/unit/google/cloud/operators/test_dataproc.py -q
125 passed, 1 warning in 97.29s

A scratch harness executed two operator instances against a local Dataproc hook and the real TaskStateStoreAccessor path:

Commands and raw crash/retry output

$ AIRFLOW_HOME="$PWD/.build/airflow-home-baseline-approved" PYTHONPATH="$PWD/.build/dataproc-baseline/airflow-core/src:$PWD/.build/dataproc-baseline/task-sdk/src:$PWD/.build/dataproc-baseline/providers/google/src" PATH="$PWD/.venv/bin:$PATH" uv run --project providers/google python dev/dataproc_resumability_e2e.py --label baseline | tee .build/dataproc-proof-baseline-approved.log
{
  "first_attempt_error": "WorkerCrash: simulated worker crash while waiting for Dataproc job",
  "label": "baseline",
  "link_job_ids": [
    "dataproc-job-1",
    "dataproc-job-2"
  ],
  "recovered_without_resubmission": false,
  "second_operator_job_id": "dataproc-job-2",
  "second_result": "dataproc-job-2",
  "status_reads": [],
  "stored_state_after_crash": {},
  "submissions": [
    "dataproc-job-1",
    "dataproc-job-2"
  ],
  "waited_job_ids": [
    "dataproc-job-1",
    "dataproc-job-2"
  ]
}
$ AIRFLOW_HOME="$PWD/.build/airflow-home-patched-approved" PYTHONPATH="$PWD/airflow-core/src:$PWD/task-sdk/src:$PWD/providers/google/src" PATH="$PWD/.venv/bin:$PATH" uv run --project providers/google python dev/dataproc_resumability_e2e.py --label patched | tee .build/dataproc-proof-patched-approved.log
{
  "first_attempt_error": "WorkerCrash: simulated worker crash while waiting for Dataproc job",
  "label": "patched",
  "link_job_ids": [
    "dataproc-job-1",
    "dataproc-job-1"
  ],
  "recovered_without_resubmission": true,
  "second_operator_job_id": "dataproc-job-1",
  "second_result": "dataproc-job-1",
  "status_reads": [
    "dataproc-job-1"
  ],
  "stored_state_after_crash": {
    "dataproc_job_id": "dataproc-job-1"
  },
  "submissions": [
    "dataproc-job-1"
  ],
  "waited_job_ids": [
    "dataproc-job-1",
    "dataproc-job-1"
  ]
}

Was generative AI tooling used to co-author this PR?

Yes - GitHub Copilot CLI (GPT-5.6 Sol)

Generated-by: GitHub Copilot CLI (GPT-5.6 Sol) following the guidelines

Synchronous Dataproc tasks can submit duplicate jobs when a worker exits after submission but before task completion. Generated-by: GitHub Copilot CLI (GPT-5.6 Sol) Signed-off-by: 1fanwang <[email protected]>

potiuk · 2026-09-25T20:22:33Z

Hello @1fanwang - thank you for your contributions to Apache Airflow!

The Airflow community has introduced a limit of 5 open pull requests at a time for contributors without write access to the repository. You currently have 34 open pull requests, so - as a one-time step of introducing the limit - we closed the ones where maintainers have not engaged yet:

Report failure causes and bound infrastructure replacements #66405 - Report failure causes and bound infrastructure replacements
Honor AirflowFailException raised inside on_retry_callback #66781 - Honor AirflowFailException raised inside on_retry_callback
Emit OpenTelemetry spans from airflow CLI entry points #66789 - Emit OpenTelemetry spans from airflow CLI entry points
Preserve sensor start_date when scheduler advances reschedule to queued #66790 - Preserve sensor start_date when scheduler advances reschedule to queued
Reuse the resumable core for TriggerDagRunOperator's durable wait #68955 - Reuse the resumable core for TriggerDagRunOperator's durable wait
Let TriggerDagRunOperator own its execution via a new accessor #69135 - Let TriggerDagRunOperator own its execution via a new accessor
Add resumable_job.reconnect_failure metric for failed reconnects #70653 - Add resumable_job.reconnect_failure metric for failed reconnects
Allow configuring the collation of asset name columns #71249 - Allow configuring the collation of asset name columns
Fix task_instance_history downgrade failing under the PyMySQL driver #71258 - Fix task_instance_history downgrade failing under the PyMySQL driver
Correct the retry_db_transaction docstring about provide_session #71311 - Correct the retry_db_transaction docstring about provide_session
Show worker pod logs for tasks that failed before writing any log #71566 - Show worker pod logs for tasks that failed before writing any log
Reconnect Synapse Spark batches on task retry #72368 - Reconnect Synapse Spark batches on task retry
Reconnect Azure Data Factory pipelines on task retry #72399 - Reconnect Azure Data Factory pipelines on task retry
Prevent duplicate Anthropic batches on task retry #72400 - Prevent duplicate Anthropic batches on task retry
Reconnect AWS Batch jobs after worker crashes #72413 - Reconnect AWS Batch jobs after worker crashes
Reconnect Airbyte sync jobs after task retries #72414 - Reconnect Airbyte sync jobs after task retries
Reconnect Data Fusion pipelines after task retries #72415 - Reconnect Data Fusion pipelines after task retries
Prevent duplicate OpenAI batches on task retry #72416 - Prevent duplicate OpenAI batches on task retry
Recover Dataproc jobs after worker crashes #72417 - Recover Dataproc jobs after worker crashes
Recover Athena queries after worker crashes #72419 - Recover Athena queries after worker crashes
Reconnect Cloud Batch jobs on task retry #72420 - Reconnect Cloud Batch jobs on task retry
Prevent duplicate ECS submissions on task retry #72421 - Prevent duplicate ECS submissions on task retry
Recover Databricks SQL statements after task retries #72422 - Recover Databricks SQL statements after task retries
Reconnect dbt Cloud runs after worker crashes #72423 - Reconnect dbt Cloud runs after worker crashes
Restore missing serialized Dags without a full reparse #72927 - Restore missing serialized Dags without a full reparse
Report task failure causes without changing retries #73350 - Report task failure causes without changing retries

These pull requests stay open because maintainers are already engaged in them - they count towards your limit:

Add durable option to TriggerDagRunOperator to reconnect on retry #68936 - Add durable option to TriggerDagRunOperator to reconnect on retry
Let deferrable tasks reconnect without polling #68952 - Let deferrable tasks reconnect without polling
Keep database failures from looking like missing secrets #71307 - Retry transient database errors in the metastore secrets backend
Return each attempt's own URL from operator extra links #71522 - Return each attempt's own URL from operator extra links
Return 404 when a DagRun disappears during task API writes #72907 - Return 404 instead of 500 when DagRun is missing in Execution API
Edge: Return 404 for missing task logs #72908 - Edge: Return 404 for missing task logs and emit failed job metrics
Edge: Return 400 instead of 500 on invalid worker version #72925 - Edge: Return 400 instead of 500 on invalid worker version
Return 404 for task-state writes after a Dag run is deleted #72926 - Return 404 for task-state writes after a Dag run is deleted

This is not a judgement of you or of your changes. We never told contributors before that opening many pull requests at once was a problem, so there is nothing to feel bad about - and nothing is lost: your branches, commits and the review history stay where they are.

What we ask you to do is to make your first prioritization decision: choose which of the pull requests above matter most to you, and reopen them (up to 5 open at a time, including the ones still open) with the "Reopen pull request" button or gh pr reopen <PR_NUMBER> --repo apache/airflow. Reopen the ones you are ready to follow through - keep them rebased, respond to review comments and fix failing checks.

While your pull requests are waiting for review, the most valuable thing you can do is help in other ways - reviewing other contributors' pull requests, helping with issues, and taking part in the discussions on the devlist and Slack.

Why we introduced the limit, what it means for you and how to reopen or restore a pull request is explained in https://github.com/apache/airflow/blob/main/contributing-docs/32_open_pull_request_limit.rst.

Drafted-by: Claude Code (Opus 5); reviewed by @potiuk before posting

Recover Dataproc jobs after worker crashes

b1c5225

Synchronous Dataproc tasks can submit duplicate jobs when a worker exits after submission but before task completion. Generated-by: GitHub Copilot CLI (GPT-5.6 Sol) Signed-off-by: 1fanwang <[email protected]>

boring-cyborg Bot added area:providers provider:google Google (including GCP) related issues labels Sep 2, 2026

potiuk added the closed because of open PR limit Closed as a one-time step of introducing the open pull request limit label Sep 25, 2026

potiuk closed this Sep 25, 2026

potiuk mentioned this pull request Sep 25, 2026

Report task failure causes without changing retries #73350

Closed

2 tasks done

Navigation Menu

Repository navigation

Recover Dataproc jobs after worker crashes - #72417

Recover Dataproc jobs after worker crashes#72417
1fanwang wants to merge 1 commit into
apache:mainfrom
1fanwang:recover-dataproc-jobs

1fanwang commented Sep 2, 2026 •

edited

Loading

Uh oh!

potiuk commented Sep 25, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

1fanwang commented Sep 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Testing

Was generative AI tooling used to co-author this PR?

Uh oh!

potiuk commented Sep 25, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

1fanwang commented Sep 2, 2026 •

edited

Loading